SCIP leaks memory

[ppershing]

SCIP solver does not release (some?) of its memory after the model is deleted. This makes it a huge pain to use Numberjack in scripts with solve a lot of problems of big size.

[ppershing]

Easy script to reproduce:

import Numberjack as numberjack
import gc


def solve():
        model = numberjack.Model()
        N = 1000
        M = 1000
        variables = [numberjack.Variable(0, 1) for i in range(N)]
        for j in range(M):
                model.add(numberjack.Sum(variables) == 42)
        solver = model.load('SCIP')
        solver.solve()

while True:
        solve()
        print gc.collect()

[9thbit]

Hi @ppershing, thanks for the bug report. I haven't been able to figure this one out yet, it appears that the memory for the variables is not freed until the program exits. I suspect this is something to do with the swig interface we use, but it requires further work. If you need a workaround for now, you could launch each solve in a separate subprocess.

Barry

[ppershing]

Yes, I already used sub-process workaround but it is a bit of a pain to smuggle data between them (so far I used multiprocessing but it requires pickleable data structures). Anyway, the issue is interesting as a different solver (I tried Gurobi) seem to work well and free the memory.